Milestone 4B: retrieval result cleanup

[princerm06]

Summary

• deduplicate semantic search results by source page while retaining the strongest matching chunk
• fetch a wider candidate pool so duplicate chunks do not reduce the number of unique sources returned
• replace raw chunk payloads with compact, whitespace-normalized excerpts
• expose clearer matched-chunk provenance fields in the search API
• simplify frontend result cards so chunk debugging metadata is no longer user-facing
• add retrieval formatting/deduplication unit tests
• extend backend CI to run both chunking and retrieval tests

Safety / scope

• no database schema changes
• no migrations
• no deployment changes
• no secrets or external services added
• main remains unchanged until this PR is verified and merged

Verification

GitHub Actions should run the backend unit-test suite automatically on this PR. Leave this PR unmerged until CI is green and the behavior is reviewed.